A cute mix that really likes making princess-type characters. This mix has a very strong style, so the checkpoint you choose may not affect things very much when you're running it at strength 1.0.
On the other hand, you'll get a very nice anime look at lower strengths (try 0.35, for example), and a dark, brooding, realistic shading look if you dip a bit into the negatives.
Description
FAQ
Comments (4)
Do you have a guide or tutorial on how you train these styles?
Dataset Selection / Captioning / Etc..
You produce really great aesthetics that are fairly flexible.
The ones with "mix" in the name are mixes. The key that I've found for creating stable LoRA and Lyco mixes is to mix a fairly large number together, at low strength. A bunch of things at strength 0.2 to 0.4 will in my experience be more stable than two things at 0.8 or 1.
As for the trained ones, if you're training in a concept, counterintuitively, it's usually best to train directly on top of the original Stable Diffusion 1.5 checkpoint. Styles are better trained on top of a more coherent checkpoint -- I usually use the AnyLora ones for that.
More to add here... For training data, I usually just go to pinterest and search there, then save a bunch of stuff, generally around 100-200 images. Sometimes I'll train something once and then discover that there's too much of a certain type of image, so I'll remove some of those. I also will sometimes mix my trained loras into something more stable, which can make them look nicer. For captions, I tend to either make simple ones myself or run them though BLIP and/or the WD 1.4 Swin v2 tagger in the automatic1111 Dataset Tag Editor extension. I avoid trigger words unless they're absolutely necessary.
The key with training LoRAS is that sometimes you need to experiment with different settings (do you train a Lyco or a Lora, how many epochs, what learning rate, etc?) and different types of captions. I'll often train mine more than once and pick the best one.
Maybe I should write an article about this. :)
@_Envy_ This was incredibly helpful..
I think if you wrote an article on how you caption datasets and your training parameters.. and even your thought process behind that and in mixing.. would all be incredibly useful..
There are many tutorials on how to do it technically.. but the finesse it seems is making good / bad decisions for producing things that are flexible and look amazing as the final outcome.
Even how you test LORA's and deduce "oh this concept is too much, let me go back and modify my dataset".
My harddrive is filling up with a lot of your stuff - you just keep producing really aesthetically pleasing and useful models and lora's.